Track UT gaps for LLVM-path vs SPIRV-path.

[tdeng5]

Issue Description

Collect the latest UT report for both LLVM-path and SPIRV-path, check the failed cases. High priority for fixing those bugs which failed in LLVM-path while pass in SPIRV-path.

Code Example

System Info

• OS:
• PyTorch:
• Intel Extension For Pytorch:
• LLVM Version.
• Triton Version:
• Intel® XPU Backend for Triton*:

[ESI-SYD]

UT Status in main branch SPIRV-path

============ Overview ============
num_language: 6859
num_others: 901
num_total: 7760
num_skipped: 1158
num_cuda_only_skipped: 637
num_failed: 995
num_passed: 5607
total_pass_rate: 77.49%
num_new_solved_expected_cases: 5
num_new_solved_todo_cases: 32
num_new_failed_cases: 943
num_new_failed_not_todo_cases: 0

failed cases

UT failed cases in llvm-target branch: llvm-path

============ Overview (core tests) ============
=== 5422 passed, 392 skipped, 1165 xfailed, 56 warnings in 816.25s (0:13:36) ===
=== 5422/(5422+392+1165) = 5422/6979 = 77.69% ===

xfail cases

test_core.py::test_atomic_rmw #152
test_core.py::test_convert2d [skip]
test_core.py::test_convertmma2mma [test_mma2mma is not supported in HIP/XPU]
test_core.py::test_dot [skip]
test_core.py::test_fp8_dot_acc [test_fp8_dot_acc not supported on XPU]
test_core.py::test_reduce_layouts #156

cuda only

test_core.py::test_globaltimer
test_core.py::test_num_threads
test_core.py::test_smid
test_core.py::test_inline_asm[1]
test_core.py::test_inline_asm_multiple_outputs
test_core.py::test_inline_asm_packed[1]
test_core.py::test_inline_asm_packed_multiple_outputs
test_core.py::test_inline_asm_with_pointers[1]

skipped as FIXME issue

test_core.py::test_scan2d #153
test_core.py::test_locality #154
test_core.py::test_scan_layouts #155
test_core.py::test_reduce_layouts #156
test_core.py::test_chain_reduce #164
test_core.py::test_masked_load_shared_memory #158
test_conversions.py::test_typeconvert_downcast #212
test_conversions.py::test_typeconvert_upcast #213

Note

No more cases passed in spirv-path but failed in llvm-path in language cases.

Non-language cases not ported and enabled in llvm-path

operator: #167 runtime: #206 tools: cuda releated

Non-language cases ported and PASSED in SPIRV-PATH (issues tracked)

operators/test_blocksparse.py::test_softmax[64-1871-False]
operators/test_blocksparse.py::test_softmax[128-2511-False]
operators/test_blocksparse.py::test_softmax[32-576-False]
operators/test_blocksparse.py::test_softmax[16-256-False]

operators/test_blocksparse.py::test_attention_fwd_bwd[dtype0-16]
operators/test_blocksparse.py::test_attention_fwd_bwd[dtype1-16]
operators/test_blocksparse.py::test_attention_fwd_bwd[dtype0-32]
operators/test_blocksparse.py::test_attention_fwd_bwd[dtype1-32]
operators/test_blocksparse.py::test_attention_fwd_bwd[dtype0-64]
operators/test_blocksparse.py::test_attention_fwd_bwd[dtype1-64]

[etiotto]

Overall I would say no surprises here. This information was known already, I have opened issues to track the functional problems which are also linked in this work item.

[ESI-SYD]

closing, find specified work item from opening issues, feel free to reopen

[vlad-penkin]

llvm branch have reached the parity with the spirv on the ut pass rate